How to extract tables from a PDF

First: which kind of PDF is it

Open it and try to select a number with your cursor. If it highlights, the characters are real. If nothing happens, the page is an image and you are in OCR territory — a different problem with a different error rate. This one check changes which of the routes below are even available, and most guides skip it.

Route 1 — copy and paste (exact, when it works)

For a text PDF, select the table and paste it. No model, no error rate: the characters transfer exactly. The mess afterwards is structural — everything lands in one column, or columns split on the wrong spaces.

Fix it with a split step rather than by hand: paste into a sheet, then split on the delimiter that actually separates the columns. Our CSV viewer will show you what you pasted; a text-to-columns step in your spreadsheet does the reshaping. Tedious, and still the most accurate route available.

Route 2 — a library that parses the PDF structure

Tools like Tabula and Camelot, or Python libraries, read the PDF's internal layout — text positions and ruling lines — and reconstruct the table. Exact on the characters, since nothing is being recognized, and good on tables with visible ruling lines. They struggle where columns are implied by whitespace only, and they need you to be comfortable running them.

Route 3 — an online converter

Fast, free, and uploads your document. Fine for a public report. Not fine for a bank statement, a payslip or anything under an NDA — and worth reading the retention policy before deciding, treating anything unstated as unbounded.

Route 4 — a model reads it

The only option for scanned pages, and the most forgiving on messy layouts, because it infers structure the way a person would. It is also the only route that can be confidently wrong: a misread decimal produces a plausible number, not an error.

Neither the tools on this site nor TableDI 2 take this route: they read PDFs that already contain text — the PDF table reader here does it in your browser — and say so when a page turns out to be a scan.

The check that makes any of this safe

Whichever route: reconcile a total against a number you already know before building anything on the result. Row count against the document, a column total against the stated total. It takes thirty seconds and it catches the decimal errors, the page-break losses and the repeated-header rows that all four routes produce.

A worked version of this for statements specifically: PDF financial statements to Excel.